Model Card for Sprakbanken/trocr_smi
This is a TrOCR-model for OCR (optical character recognition) of Sámi languages.
It can be used to recognize text in images of printed text (scanned books, magazines, etc.) in North Sámi, South Sámi, Lule Sámi, and Inari Sámi.
How to Get Started with the Model
Use the code below to get started with the model.
1from transformers import TrOCRProcessor, VisionEncoderDecoderModel
2from PIL import Image
3
4processor = TrOCRProcessor.from_pretrained("Sprakbanken/trocr_smi")
5model = VisionEncoderDecoderModel.from_pretrained("Sprakbanken/trocr_smi")
6
7image = Image.open("path_to_image.jpg").convert("RGB")
8
9pixel_values = processor(image, return_tensors="pt").pixel_values
10generated_ids = model.generate(pixel_values)
11
12generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
Model Details
This model is
microsoft/trocr-base-printed fine-tuned on manually annotated Sámi OCR data.
See our paper for more details.
Model Description
- Developed by: The National Library of Norway
- Model type: TrOCR
- Languages: North Sámi (sme), South Sámi (sma), Lule Sámi (smj), and Inari Sámi (smn)
- License: CC BY 4.0
- Finetuned from model : microsoft/trocr-base-printed
Model Sources
- Repository: https://github.com/Sprakbanken/nodalida25_sami_ocr
- Paper: "Enstad T, Trosterud T, Røsok MI, Beyer Y, Roald M. Comparative analysis of optical character recognition methods for Sámi texts from the National Library of Norway. Accepted for publication in Proceedings of the 25th Nordic Conference on Computational Linguistics (NoDaLiDa) 2025." preprint
Collection details
This model is a part of our collection of OCR models for Sámi languages.
The following TrOCR models are available:
- Sprakbanken/trocr_smi: microsoft/trocr-base-printed fine-tuned on manually annotated Sámi data
- Sprakbanken/trocr_smi_nor: microsoft/trocr-base-printed fine-tuned on manually annotated Sámi and Norwegian data
- Sprakbanken/trocr_smi_pred: microsoft/trocr-base-printed fine-tuned on manually annotated and automatically transcribed Sámi data
- Sprakbanken/trocr_smi_nor_pred: microsoft/trocr-base-printed fine-tuned on manually annotated and automatically transcribed Sámi data, and manually annotated Norwegian data
- Sprakbanken/trocr_smi_synth: microsoft/trocr-base-printed fine-tuned on Sprakbanken/synthetic_sami_ocr_data, and then on manually annotated Sámi data
- Sprakbanken/trocr_smi_pred_synth: microsoft/trocr-base-printed fine-tuned on Sprakbanken/synthetic_sami_ocr_data, and then fine-tuned on manually annotated and automatically transcribed Sámi data
- Sprakbanken/trocr_smi_nor_pred_synth: microsoft/trocr-base-printed fine-tuned on Sprakbanken/synthetic_sami_ocr_data, and then fine-tuned on manually annotated and automatically transcribed Sámi data, and manually annotated Norwegian
Sprakbanken/trocr_smi_pred_synth is the model that achieved the best results (of the TrOCR models) on our test dataset.
Uses
You can use the raw model for optical character recognition (OCR) on single text-line images in North Sámi, South Sámi, Lule Sámi, and Inari Sámi.
Out-of-Scope Use
The model only works with images of lines of text.
If you have images of entire pages of text, you must segment the text into lines first to benefit from this model.
Citation
APA:
Enstad, T., Trosterud, T., Røsok, M. I., Beyer, Y., & Roald, M. (2025). Comparative analysis of optical character recognition methods for Sámi texts from the National Library of Norway. Proceedings of the 25th Nordic Conference on Computational Linguistics (NoDaLiDa).
